EmoPerrezDataAnalysis

Hypothesis Testing

• Comparing the results with hypothesis and human
• Here is the list of questions:

H1:

1. Effects of situation and process characteristics on subjective appraisals and on emotional reactions of a stressful episode:
   • 1.1) According to the objectives of the construction, subjects should perceive higher controllability and changeability in the aversive episdoes than in the loss or failure episodes
   • 1.2) In both types of stressful situations controllability and changeability should be perceived as lower over the course of the episodes, from phase 1 to phase 2, due to the stressor remaining unchanged over time.
   • 1.3) Negative valence should be similarly perceived in the two subtests (aver vs loss), but should be in the course of the (unchanged) stressful episdoes. There should be a strong difference between positive and negative outcome as described in the hypothetical end of the episode (phase 3)
   • 1.4) Concerning the theme or content of the two classes of stressful episdoes, it is to be supposed that aversive stimuluation by social agents (as presented in some episodes) produces more aggresive and less depressed emotions** than the looming loss or failure events

H2:

2. Effects of situation and process on situation-directed behaviors:
   • 2.1) The less controllable the situations (gross and net control), the less active influence on the stressor will be exerted by the subject and,
   • 2.2) the more passivity (hesitation and resignation) will occur.

H3:

3. Effects on representation-directed coping efforts: The lower the ambiguity of the stressful situation then
   • 3.1) the more limited the serach for information, and
   • 3.2) the more the subject will surppress information about the stressor.

H4:

4. Evaluation-directed coping efforts: The less controllable the situation,
   • 4.1) The more the instrumental goals (of influence, substituation) are changed from more alloplastic to more autoplastic adaptation, especially when stressful effects are long-lasting; because social aversive stressors as convceived in the SCPQ episodes probably represent a stronger attack on the self-esteem of the subject, his or her coping intentions should focus more on the goal of (re)equilibration of self-esteem;
   • 4.2) the more re-evaluation is increased to change the relavance or valence of the actual situation.

Results appearing on the main text

• Combine all results into one image

Act as a person with depression

Both types

Aversive stories only

Loss&Failure stories only

Results appearing in the main text (Only Aversive)

Within Models Across Setups

• Setup: Individual/Batch (2) x Before/After (2) x Choice/Num (2) = 8
• Two views: between setups (testing hypo) and within setup (comparing similarity)

Within Setup

Results appearing in the main text

Between setups

Results in the main text

Model#1: Text-Davinci-003

Model#2: Chat

Model#3: GPT4

Conditioned on previous appraisal answers

• Reported this in the main paper but didn’t show any figure.

Hypotheses Testing Results

• A Heatmap/table shows the number of setups for a given model align with hypotheses for different theshold values
• Regression is fitted in a similar as the Perrez work, specifically fitting the interaction effect between type and time.
• We decide to not report these in the main text because of the following reasons:
• First, if we were to report any results based on any cutoff threshold, this wouldn’t be fair to the model due ot the fact that they only have 9 data point per condition. In addition, one could argue that with temperature = 0, you are always going to get the same results. Therefore, you can collect 100 more data and the mean would be the same but SE would limit to 0.
• Second, hence, one may argue that then looking at the sign/direction would be enough. However, we feel that this would also going to bias the view since presenting with only sign results ignore magnitude which clearly differs from human.
• Therefore, we decide to not inlcude the following results in the main paper.
• To address this problem, the future work will look at temperature !=0 and repeat multiple times (require time and money)
• Another way to address this problem is to get new human data so we can compare magtinude of human and models.

## Using question as id variables

## Using question as id variables

## Using question as id variables

## Using question as id variables